- Mar 11, 2000
- 23,586
- 1,000
- 126
It's really the GPU number that throws me off at 8.3%, CPU is fine at 16%. Scaling GPU with more transistors is straightforward, but it seems like they didn't do that this time and this was just an iterative improvement and keeps the overall 4-core design from A13 with potentially higher clocks to take advantage of the new process node.
Fair! More worried about the 8% GPU increase, which sounds like "we put the A13 GPU on N5." But I'm getting talked into the idea that they moved graphics engineers onto another pipeline, and we'll see a new GPU core debut with a Mac or iPad Pro.
I'm more interested in power efficiency numbers.
Marvell getting 40% powersavings for a chip that consumes 0.6x total from a similar 7nm SoC.
Alchips getting 64% powersavings for a chip that consumes 0.36x total from a similar 7nm SoC.
Didn't the A13 have thermal issues? Where it was getting beat by Kyro/Adreno sometimes do to it.
1. Reduced voltage for same drive current => High Mobility, etc
2. Reduced resistance/capacitance => Improved Low-K, RIE, Graphene cap, Via pillar, etc
3. Reduced area penalty/increased routing efficiency => COAG, EUV, etc
Marvell getting 40% powersavings for a chip that consumes 0.6x total from a similar 7nm SoC.
Alchips getting 64% powersavings for a chip that consumes 0.36x total from a similar 7nm SoC.
Didn't the A13 have thermal issues? Where it was getting beat by Kyro/Adreno sometimes do to it.
1. Reduced voltage for same drive current => High Mobility, etc
2. Reduced resistance/capacitance => Improved Low-K, RIE, Graphene cap, Via pillar, etc
3. Reduced area penalty/increased routing efficiency => COAG, EUV, etc
Last edited:
This is an A14 and not an A14X. Putting more GPU units instead of increasing frequency or going to a new architecture means extra die space and extra die space means less iPhone and iPad Air chips due to TSMC 5NM being the latest process which is brand new.
It is like Tick Tok Philosophy of Intel when Intel was good 7 to 10 years ago. You see more ambitious GPU and other stuff during the refresh when you can do more die space for it is not the leading edge. First new chip though for a new process is going to be the smaller version, the year after you do the more ambitious design for the same process.
I can't prove this (but I bet it is true) that we did not see a new iPad Pro chip for it will be using an A14x likely the same chip in the upcoming ARM / Apple Silicon Macbooks / Macbook Pros. Those chips will have a higher die size than the A14 and thus Apple is prioritizing iPhone sales right now for 2 to 3 months and having iPad Pros and Apple Silicon Macs ship at the end of this year but really it is for next quarter numbers like they did with the MacPros of years past.
Sidenote last year everything that was cheap (well almost everything) was shipping with A10s aka a 10nm processor, by moving to A12s with the cheap iPad now that is a 7nm processor with the best iPhone and the iPad Airs now on 5nm.
Homepods from 2 years ago use A8 (if I recall this is TSMC 20nm) and 1GB of ram. I heard speculations the Homepods may get an update soon.
------
Side question for comparison sake has anyone figured out how many transistors are in the new TigerLake chips from Intel?
Last edited:
- Mar 11, 2000
- 23,586
- 1,000
- 126
Regarding future SoC minimums for OS support:
Truthfully I think only iPod touch is truly relevant here since it’s the only one of the three devices mentioned still being sold with an SoC lower than A12 that actually runs iOS proper. But even then it’s kinda iffy because Apple seems to consider iPod touch a second rate iDevice, at least compared to iPhones and iPads.
Regardless though, while A12 will become the minimum in time, I don’t think it will be anytime soon. There are just too many recent A10 devices out there. iPad 6 and iPad 7 are both A10. Furthermore, even A8 is still getting support in iPadOS 14. They’re not going to jump from A8 all the way up to A12 in one go.
The next couple of years of iPadOS and iOS after 14 will be with A10 minimum and then after that A12 will be the new baseline.
Truthfully I think only iPod touch is truly relevant here since it’s the only one of the three devices mentioned still being sold with an SoC lower than A12 that actually runs iOS proper. But even then it’s kinda iffy because Apple seems to consider iPod touch a second rate iDevice, at least compared to iPhones and iPads.
Regardless though, while A12 will become the minimum in time, I don’t think it will be anytime soon. There are just too many recent A10 devices out there. iPad 6 and iPad 7 are both A10. Furthermore, even A8 is still getting support in iPadOS 14. They’re not going to jump from A8 all the way up to A12 in one go.
The next couple of years of iPadOS and iOS after 14 will be with A10 minimum and then after that A12 will be the new baseline.
Well it depends on what Apple is targeting, and what gets targeted. As pointed out, this is the regular A14 that will go in the phone. Does the phone need a big increase in GPU performance this year? It sounds like they felt the biggest increase was needed in NPU performance. This is designed for the phone, the iPad Air is not the high end tablet so it gets what the phone needs not the other way around.
Based on the performance increase claimed for the NPU, it would appear they doubled the number of transistors devoted to it. Maybe they need that to begin their AR push, or other initiatives that will become clear later.
The A14X (and whatever goes into the soon to be released Macs, if it is separate from the A14X) will obviously have additional transistors beyond the two additional big cores. If the goals for the products it goes into are different than the goals of the phone, it may see more area dedicated to the GPU and a resulting performance increase.
I'm looking forward to AT's eventual deep dive on this chip. Raw performance gains may not be impressive, but that could always be a case of greater focus on efficiency.
Considering these are SoCs for mobile devices that's just as important as the raw power capabilities. Apple blows everyone else out of the water when it comes to the raw performance anyhow so it isn't as though they're getting left behind if the actual uplift is low double-digits.
Considering these are SoCs for mobile devices that's just as important as the raw power capabilities. Apple blows everyone else out of the water when it comes to the raw performance anyhow so it isn't as though they're getting left behind if the actual uplift is low double-digits.
- Mar 11, 2000
- 23,586
- 1,000
- 126
DigiTimes is claiming that Apple has filled up all of TMSC's 5 nm capacity.
I can't provide a link because it hasn't gone to online press yet.
I can't provide a link because it hasn't gone to online press yet.
Indeed. I've been saying that A14 is more than fast enough to run a MacBook non-Pro, but others have rightly argued that from a marketing point of view that could prove more difficult and the fact is that they don't need to go with A14 from a power utilization perspective. An A14X or A14X variant (either at the same speed as the iPad Pro or else maybe even downclocked) would mean a very big increase in MacBook performance while at about the same power usage. A revival of the 12" MacBook line with fanless A14X is feasible, even if A14 non-X is fast enough.
I don't think there's any doubt that the A14X will be larger and devote more die area to the GPU. We've already seen Apple double GPU cores with the A12X.
What I'm wondering about is more to do with whether or not the cores themselves will see changes. As you say, the phone is already well-positioned with its GPU. An 8% boost from the node shrink still leaves it well ahead of rivals, and more than capable for what the iPhone does. But if Apple scales this core up the way they did going from the A12 to the A12X, using essentially the same core but doubling the number, does it provide an excellent level of performance for notebook machines?
There were some dodgy rumors recently about Apple developing a GPU for 2021. I'm not sure I buy an Apple dGPU next year, but I'm a lot more willing to buy the idea that they have a different GPU core design that would show up in APUs next year with the same A14 perf cores, ahead of the A15's launch.
By everyone else I assume you mean “in the mobile phone and ARM based tablet space” because there are dozens of chips out there that walk all over Apple’s “raw performance”.
I'm wondering if that's exactly it: essentially no changes made to either CPU or GPU...
The watch has now had two rounds of staggered CPU.
aWatch5 was based on same core as aWatch4, the A12 small core;
aWatch6 is now based on the A13 small core.
Likewise the X processors have adopted a two year cadence.
Maybe with the Apple Silicon Macs, Apple just hit the wall in terms of how much can be renewed every year?
So maybe from now on the pattern is something like a core that's more appropriate for the high end in even years, and a middleweight core for iPhones in odd years?
That's somewhat simplifying it -- there are details like how the small core gets updated each year, and obviously each core will share DNA, more like Xeon vs Core rather than Xeon vs Atom --- but that basic idea?
So maybe the reason Apple said so little about the core/gpu is that there is very little to say!? It's good old A13! They took a ~15% free iso-power frequency boost from N5 to goose the core, did not take that frequency boost for the GPU (meaning it will be the same speed as the old GPU at peak, but, N5..., should be able to maintain that peak a lot longer maybe indefinitely).
Meanwhile perhaps the A14X/Z/Ω had the CPU team working on something without the compromises inherent in a phone design. Power budget running from? 12W to 60W? A lot of IO connections? Some sort of GPU story -- maybe the fabled chiplets?
Maybe the end result of all this is that iPad Pro and Apple Silicon ship, not as we expected, with an A14 variant at all, but with a new K (or N or R or whatever) series chip?
We always expected the cores to be shared, and the GPU was an unknown problem. But we've seen the trouble Intel has got itself into trying to scale a single core from Xeons down to phones. Maybe Apple concluded
(a) that was just a dumb idea. Share as much as you can of the design, but they can't be identical.
(b) there are enough other things that can be changed in an iPhone every year, both on the board and on the SoC, that CPU and GPU can be updated every two years.
I miss Anand and Brian doing podcasts about Silicon. Ian is great but I still adore my two originals.
Fab18-P1 and Fab18-P2 look pretty much up and running via google maps.
Basically, there is two 40K+ fabs running right now. With that third one running in 2021 making N5 in larger quantity than N10/N7/N6 combined.
Q1-2020 => Apple orders 45,000 wafers for all three months => 15,000 wafers top per month // 25K wpm (w/o P2) left over
Q4-2020 => Apple orders 5,000-6,000 wafers for one month => 18,000 wafers top for all three months. // 220K wpq left over
¯\_(ツ)_/¯
If anything Apple is down-ramping as it has everything already.
Last edited:
Yeah the bandwidth of their teams, and the timelines for product introductions have to be considered. If the CPU team's timeline is "has to be ready to ship by September 2020" and the GPU team's timeline is "has to be ready to ship by December 2020" they might have planned all along to put the A13 GPU core in the A14 - if they gained 8% clock rate from the N7P -> N5 shrink that's all they'd get.
Future iterations of the A14 for iPad Pro and the Mac line get the new GPU core that has more improvements, and those are the products that can both benefit more from an improved GPU and are able to wait a few more months to get it.
- Mar 11, 2000
- 23,586
- 1,000
- 126
Are you trying to say Apple has stockpiled enough A14 chips in Q1 2020? That doesn’t seem right to me.
I realize that DigiTimes isn’t the most reliable source, but how does a satellite photo of a building exterior indicate that they have 5 nm production ramped up and at good yields?
Based on 125 mm2 die being worst-case. Q1-Q3 is about ~18+ million devices. There is a down ramp not a up ramp for Q4 2020. So, the indication is that they enough of what they need to down ramp. Even then, they have never used TSMC 5nm full capacity.
Yields have been good since 2019. - TSMC announced it with a 512 Mb Cache and CPU/GPU/etc test chip.
The satellite proves that Phase 2 is finished and getting power hence the smoke.
All phases for Fab 18 have a 6 month period from being equipped and ready to go full HVM (>40,000 wafer starts per month per phase at +2Q each)
Last edited:
- Mar 11, 2000
- 23,586
- 1,000
- 126
- Mar 11, 2000
- 23,586
- 1,000
- 126
Article from Q2 2020:
hexus.net
A report published by Taiwan's Commercial Times newspaper provides an interesting glance at the future, with a list of TSMC N5 and N5+ customers plus their respective chip designs. TSMC's N5 node will begin mass production starting in Q3 this year and the Commercial Times provides a timetable of node allocation starting later this year with orders from Apple and Huawei being fulfilled first.
All testing so far has indicated that the only supported YouTube formats in iOS 14 / iPadOS 14 and macOS Big Sur in Safari are h.264 and VP9.
I suppose Apple could build hardware AV1 decode into the SoC and just choose not to expose it until later, but I might wonder if they wouldn't bother adding it until say A16 or something. I think from Apple's perspective there isn't really any need to actively support AV1 until about 2023-2024, so if they start shipping A16 devices with it in 2022, that will allow them to start supporting it in 2023 to include both the new A17 devices and the older A16 devices, while leaving A15 and older iDevices in the dust.
Alternatively, perhaps they could leave it off chip and add it to T3 to add AV1 decode support for the Macs first... that is if the Arm Macs even have such a chip.
List of TSMC 5nm customers and their orders published
AMD Zen 4 and RDNA 3 chips, Nvidia Hopper GPUs, Apple A15 mass production in 2021-22.
hexus.net
A report published by Taiwan's Commercial Times newspaper provides an interesting glance at the future, with a list of TSMC N5 and N5+ customers plus their respective chip designs. TSMC's N5 node will begin mass production starting in Q3 this year and the Commercial Times provides a timetable of node allocation starting later this year with orders from Apple and Huawei being fulfilled first.
I'm guessing but I consider that unlikely, considering that Apple has only started including VP9 support this year.
All testing so far has indicated that the only supported YouTube formats in iOS 14 / iPadOS 14 and macOS Big Sur in Safari are h.264 and VP9.
I suppose Apple could build hardware AV1 decode into the SoC and just choose not to expose it until later, but I might wonder if they wouldn't bother adding it until say A16 or something. I think from Apple's perspective there isn't really any need to actively support AV1 until about 2023-2024, so if they start shipping A16 devices with it in 2022, that will allow them to start supporting it in 2023 to include both the new A17 devices and the older A16 devices, while leaving A15 and older iDevices in the dust.
Alternatively, perhaps they could leave it off chip and add it to T3 to add AV1 decode support for the Macs first... that is if the Arm Macs even have such a chip.
Last edited:
Q1 2020 Taiwan Semiconductor Manufacturing Co Ltd Earnings Call:
www.tsmc.com
"N5 is already in volume production with good yield." - APRIL 16, 2020
TSMC 2022 Q2 Quarterly Results - Taiwan Semiconductor Manufacturing Company Limited
2022 Q2 Quarterly Results quarterly financial statements, presentation material, management report, earnings release earnings conference transcript. For more details regarding, please refer to TSMC IR website.
- Mar 11, 2000
- 23,586
- 1,000
- 126
Yes, and I think they chose their wording carefully, as "volume production" is not necessarily the same thing as "mass production" at iPhone SoC levels. Furthermore, even disregarding that caveat, that is in Q2 already, not the Q1 you were claiming.Q1 2020 Taiwan Semiconductor Manufacturing Co Ltd Earnings Call:
"N5 is already in volume production with good yield." - APRIL 16, 2020TSMC 2022 Q2 Quarterly Results - Taiwan Semiconductor Manufacturing Company Limited
2022 Q2 Quarterly Results quarterly financial statements, presentation material, management report, earnings release earnings conference transcript. For more details regarding, please refer to TSMC IR website.
Maybe I missed it, but AFAIK just about nobody was claiming that TSMC was in high volume mass N5 production in Q1 2020, not even TSMC themselves.
Is it even meaningful to compare performance for traditional apps anymore when so much of the transistor budget is dedicated to accelerating non-CPU and non-GPU applications?
This chip has almost has as many transistors as the RTX 2080 for crying out loud. And it has more transistors than a 3950x chip. It’s got accelerators for so many types of applications including ML, audio, camera, video editing, storage, and security. When we just look at performance for normal applications, we might miss huge increases in other areas.
Apple has data on what apps and functions people like to use and they optimize the performance of their chips for those workloads. They can de prioritize raw CPU or raw GPU performance if they see from the data that it’s not going to make that much of a difference. This is the beauty of designing your entire stack. You can optimize for exactly what your users need instead of buying a chip from a 3rd party that optimizes for many general things.
So when people on these forums make claims like Apple’s Mac chips won’t be able to touch Zen2 or Zen3 chips, they’re missing the forest for the trees. These Apple SoCs pack so many different useful accelerators that Intel and AMD processors just don’t have and can’t have because their designs optimize for the general workload. An example is the Apple T2 MAC chip which accelerates some video editing tasks that no Ryzen or Intel chip can match.
People are underestimating how much success Apple will have with ARM Macs.
This chip has almost has as many transistors as the RTX 2080 for crying out loud. And it has more transistors than a 3950x chip. It’s got accelerators for so many types of applications including ML, audio, camera, video editing, storage, and security. When we just look at performance for normal applications, we might miss huge increases in other areas.
Apple has data on what apps and functions people like to use and they optimize the performance of their chips for those workloads. They can de prioritize raw CPU or raw GPU performance if they see from the data that it’s not going to make that much of a difference. This is the beauty of designing your entire stack. You can optimize for exactly what your users need instead of buying a chip from a 3rd party that optimizes for many general things.
So when people on these forums make claims like Apple’s Mac chips won’t be able to touch Zen2 or Zen3 chips, they’re missing the forest for the trees. These Apple SoCs pack so many different useful accelerators that Intel and AMD processors just don’t have and can’t have because their designs optimize for the general workload. An example is the Apple T2 MAC chip which accelerates some video editing tasks that no Ryzen or Intel chip can match.
People are underestimating how much success Apple will have with ARM Macs.
Last edited:
TRENDING THREADS
-
Discussion Intel current and future Lakes & Rapids thread- Started by TheF34RChannel
- Replies: 23K
-
Discussion Zen 5 Speculation (EPYC Turin and Strix Point/Granite Ridge - Ryzen 9000)- Started by DisEnchantment
- Replies: 10K
-
Discussion Intel Meteor, Arrow, Lunar & Panther Lakes Discussion Threads
- Started by Tigerick
- Replies: 7K
-
-
Facebook
Twitter